mplu: Combine MPArray construction and factorization

mplu(A::AbstractArray{TW,2}; TF=nothing, TR=nothing, onthefly=nothing) where TW <: Real

Combines the constructor of the multiprecision array with the factorization.

Step 1: build the MPArray.

  (a) Store A and b in precision TW. 

  (b) Store the factorization (copy of A) in precision TF

  (c) Preallocate storage for the residual and a local copy of the
  solution in precision TR

Step 2: factor the low precision copy and return the factorization object

The TR kwarg is the residual precision. Leave this alone unless you know what you are doing. The default is nothing which tells the solver to set TR = TW. If you set TR it must be a higher precision than TW and you are essentially solving TR.(A) x = TR.(b) with IR with the factorization in TF. The MPArray structure stores the solution and the residual in precision TR and so the residual computation is done via TR.(A) x. The interprecision transfers are on the fly. So, the storage cost is the matrix, and the copy in the factorization precision.

The classic case is TW = TF = Float32 and TR = Float64. The nasty part of this is that you must store TWO copies of the matrix. One for the residual computation and the other to overwrite with the factors. I do not think this is a good deal unless A is seriously ill-conditioned. My support for this is through mplu. To do this you must put the TR kwarg explicitly in your call to mplu.

Example

julia> using MultiPrecisionArrays.Examples

julia> n=31; alpha=Float32(1.0);

julia> G=Gmat(n, Float32);

julia> A = I + alpha*G;

julia> b = A*ones(Float32,n);

# use mpa with TF=TW=Float32 and TR=Float64

julia> AF = mplu(A; TF=Float32, TR=Float64, onthefly=true);

# Solve and save the iteration history

julia> mout = \(AF, b; reporting=true);
julia> mout.rhist
4-element Vector{Float64}:
 1.12500e+00
 1.17299e-07
 1.42109e-14
 4.44089e-16

# What does this mean. I'll solve the promoted problem. TR.(A) x = b

julia> AD=Float64.(A);

julia> xd = AD\b;

julia> norm(xd - mout.sol,Inf)
8.88178e-16

# So IR with TR > TW solves a promoted problem.